Haicheng Wu
Haicheng Wu
Can you paste the error?
The error is saying that `LinearCombinationBiasRelu` does not have `operator()` that only takes one parameter which is true. `LinearCombinationBiasRelu` requires multiple sources and outputs. You need to use this epilogue:...
If you use `DefaultConv2dFpropWithBroadcast`, the ctor of `Arguments` is like https://github.com/NVIDIA/cutlass/blob/master/include/cutlass/conv/kernel/implicit_gemm_convolution_with_fused_epilogue.h#L178-L190 . Now you do not provide `ptr_Vector` and `ptr_Tensor`. `ptr_tensor` is to load an additional vector. Is that what...
then you just need to use https://github.com/NVIDIA/cutlass/blob/master/include/cutlass/numeric_conversion.h#L66-L70 with `DefaultConv2dFprop`
Sorry, I pasted wrong links. On top of https://github.com/NVIDIA/cutlass/issues/509#issuecomment-1157222714 , change `LinearCombinationBiasRelu` to `LinearCombinationRelu`
https://github.com/NVIDIA/cutlass/blob/master/tools/util/include/cutlass/util/device_nchw_to_nhwc.h#L117-L122 you code looks wrong to me. the input and output orders are wrong.
@jwang323
beta0 has to be 0 in the current implementation.
It is not going to change in 2.9
We haven't implemented it.